Implement support for WebAssembly threads

... and add a parallel raytracing demo!

This commit adds enough support to `wasm-bindgen` to produce a workable
wasm binary *today* with the experimental WebAssembly threads support
implemented in Firefox Nightly. I've tried to comment what's going on in
the commits and such, but at a high level the changes made here are:

* A new transformation, living in a new `wasm-bindgen-threads-xform`
  crate, prepares a wasm module for parallel execution. This performs a
  number of mundane tasks which I hope to detail in a blog post later on.
* The `--no-modules` output is enhanced with more support for when
  shared memory is enabled, allowing passing in the module/memory to
  initialize the wasm instance on multiple threads (sharing both module
  and memory).
* The `wasm-bindgen` crate now offers the ability, in `--no-modules`
  mode, to get a handle on the `WebAssembly.Module` instance.
* The example itself requires Xargo to recompile the standard library
  with atomics and an experimental feature enabled. Afterwards it
  experimentally also enables threading support in wasm-bindgen.

I've also added hopefully enough CI support to compile this example in a
builder so we can upload it and poke around live online. I hope to
detail more about the technical details here in a blog post soon as
well!
This commit is contained in:
Alex Crichton
2018-10-04 20:00:23 -07:00
parent 58fb907baa
commit 25b26f41e7
19 changed files with 1682 additions and 24 deletions

3
examples/raytrace-parallel/.gitignore vendored Normal file
View File

@ -0,0 +1,3 @@
package-lock.json
raytrace_parallel.js
raytrace_parallel_bg.wasm

View File

@ -0,0 +1,29 @@
[package]
name = "raytrace-parallel"
version = "0.1.0"
authors = ["The wasm-bindgen Developers"]
[lib]
crate-type = ["cdylib"]
[dependencies]
console_error_panic_hook = "0.1"
futures = "0.1"
js-sys = { path = '../../crates/js-sys' }
raytracer = { git = 'https://github.com/alexcrichton/raytracer', branch = 'update-deps' }
wasm-bindgen = { path = "../..", features = ['serde-serialize'] }
wasm-bindgen-futures = { path = '../../crates/futures' }
[dependencies.web-sys]
path = '../../crates/web-sys'
features = [
'CanvasRenderingContext2d',
'ErrorEvent',
'Event',
'ImageData',
'Navigator',
'Window',
'Worker',
'DedicatedWorkerGlobalScope',
'MessageEvent',
]

View File

@ -0,0 +1,17 @@
# Parallel Raytracing
[View documentation for this example online][dox] or [View compiled example
online][compiled]
[dox]: https://rustwasm.github.io/wasm-bindgen/examples/raytrace.html
[compiled]: https://rustwasm.github.io/wasm-bindgen/exbuild/raytrace/
You can build the example locally with:
```
$ ./build.sh
```
(or running the commands on Windows manually)
and then visiting http://localhost:8080 in a browser should run the example!

View File

@ -0,0 +1,3 @@
[dependencies.std]
stage = 0
features = ['wasm-bindgen-threads']

View File

@ -0,0 +1,25 @@
#!/bin/sh
set -ex
# Two critical steps are required here to get this working:
#
# * First, the Rust standard library needs to be compiled. The default version
# is not compatible with atomics so we need to compile a version, with xargo,
# that is compatible.
#
# * Next we need to compile everything with the `atomics` feature enabled,
# ensuring that LLVM will generate atomic instructions and such.
RUSTFLAGS='-C target-feature=+atomics' \
rustup run nightly xargo build --target wasm32-unknown-unknown --release
# Threading support is disabled by default in wasm-bindgen, so use an env var
# here to turn it on for our bindings generation. Also note that webpack isn't
# currently compatible with atomics, so we go with the --no-modules output.
WASM_BINDGEN_THREADS=1 \
cargo +nightly run --manifest-path ../../crates/cli/Cargo.toml \
--bin wasm-bindgen -- \
../../target/wasm32-unknown-unknown/release/raytrace_parallel.wasm --out-dir . \
--no-modules
python3 -m http.server

View File

@ -0,0 +1,226 @@
<html>
<head>
<meta content="text/html;charset=utf-8" http-equiv="Content-Type"/>
<style>
#scene {
height: 100%;
width: 500px;
float: left;
}
#render, .concurrency, .timing {
padding: 20px;
margin: 20px;
float: left;
}
</style>
</head>
<body>
<textarea id='scene'>
{
"width": 800,
"height": 800,
"fov": 90.0,
"shadow_bias": 1e-13,
"max_recursion_depth": 20,
"elements": [
{
"Sphere" : {
"center": {
"x": 0.0,
"y": 0.0,
"z": -5.0
},
"radius": 1.0,
"material": {
"coloration" : {
"Color": {
"red": 0.2,
"green": 1.0,
"blue": 0.2
}
},
"albedo": 0.18,
"surface": {
"Reflective": {
"reflectivity": 0.7
}
}
}
}
},
{
"Sphere" : {
"center": {
"x": -3.0,
"y": 1.0,
"z": -6.0
},
"radius": 2.0,
"material": {
"coloration": {
"Color": {
"red": 1.0,
"green": 1.0,
"blue": 1.0
}
},
"albedo": 0.58,
"surface": "Diffuse"
}
}
},
{
"Sphere": {
"center": {
"x": 2.0,
"y": 1.0,
"z": -4.0
},
"radius": 1.5,
"material": {
"coloration": {
"Color": {
"red": 1.0,
"green": 1.0,
"blue": 1.0
}
},
"albedo": 0.18,
"surface": {
"Refractive": {
"index": 1.5,
"transparency": 1.0
}
}
}
}
},
{
"Plane": {
"origin": {
"x": 0.0,
"y": -2.0,
"z": -5.0
},
"normal": {
"x": 0.0,
"y": -1.0,
"z": 0.0
},
"material": {
"coloration": {
"Color": {
"red": 1.0,
"green": 1.0,
"blue": 1.0
}
},
"albedo": 0.18,
"surface": {
"Reflective": {
"reflectivity": 0.5
}
}
}
}
},
{
"Plane": {
"origin": {
"x": 0.0,
"y": 0.0,
"z": -20.0
},
"normal": {
"x": 0.0,
"y": 0.0,
"z": -1.0
},
"material": {
"coloration": {
"Color": {
"red": 0.2,
"green": 0.3,
"blue": 1.0
}
},
"albedo": 0.38,
"surface": "Diffuse"
}
}
}
],
"lights": [
{
"Spherical": {
"position": {
"x": -2.0,
"y": 10.0,
"z": -3.0
},
"color": {
"red": 0.3,
"green": 0.8,
"blue": 0.3
},
"intensity": 10000.0
}
},
{
"Spherical": {
"position": {
"x": 0.25,
"y": 0.0,
"z": -2.0
},
"color": {
"red": 0.8,
"green": 0.3,
"blue": 0.3
},
"intensity": 250.0
}
},
{
"Directional": {
"direction": {
"x": 0.0,
"y": 0.0,
"z": -1.0
},
"color": {
"red": 1.0,
"green": 1.0,
"blue": 1.0
},
"intensity": 0.0
}
}
]
}
</textarea>
<button disabled id='render'>Loading wasm...</button>
<div class='concurrency'>
<p id='concurrency-amt'>Concurrency: 1</p>
<br/>
<input disabled type="range" id="concurrency" min="0" max="1" />
</div>
<div id='timing'>
Render duration:
<p id='timing-val'></p>
</div>
<canvas id='canvas'></canvas>
<script>
delete WebAssembly.instantiateStreaming;
document.getElementById('render').disabled = true;
document.getElementById('concurrency').disabled = true;
</script>
<script src='raytrace_parallel.js'></script>
<script src='index.js'></script>
</body>
</html>

View File

@ -0,0 +1,119 @@
const button = document.getElementById('render');
const canvas = document.getElementById('canvas');
const scene = document.getElementById('scene');
const concurrency = document.getElementById('concurrency');
const concurrencyAmt = document.getElementById('concurrency-amt');
const timing = document.getElementById('timing');
const timingVal = document.getElementById('timing-val');
const ctx = canvas.getContext('2d');
button.disabled = true;
concurrency.disabled = true;
// First up, but try to do feature detection to provide better error messages
function loadWasm() {
if (typeof SharedArrayBuffer !== 'function') {
alert('this browser does not have SharedArrayBuffer support enabled');
return
}
// Test for bulk memory operations with passive data segments
// (module (memory 1) (data passive ""))
const buf = new Uint8Array([0x00, 0x61, 0x73, 0x6d, 0x01, 0x00, 0x00, 0x00,
0x05, 0x03, 0x01, 0x00, 0x01, 0x0b, 0x03, 0x01, 0x01, 0x00]);
if (!WebAssembly.validate(buf)) {
alert('this browser does not support passive wasm memory, demo does not work');
return
}
wasm_bindgen('./raytrace_parallel_bg.wasm')
.then(run)
.catch(console.error);
}
loadWasm();
const { Scene, WorkerPool } = wasm_bindgen;
function run() {
// The maximal concurrency of our web worker pool is `hardwareConcurrency`,
// so set that up here and this ideally is the only location we create web
// workers.
pool = new WorkerPool(navigator.hardwareConcurrency);
// Configure various buttons and such.
button.onclick = function() {
console.time('render');
let json;
try {
json = JSON.parse(scene.value);
} catch(e) {
alert(`invalid json: ${e}`);
return
}
canvas.width = json.width;
canvas.height = json.height;
render(new Scene(json));
};
button.innerText = 'Render!';
button.disabled = false;
concurrency.oninput = function() {
concurrencyAmt.innerText = 'Concurrency: ' + concurrency.value;
};
concurrency.min = 1;
concurrency.step = 1;
concurrency.max = navigator.hardwareConcurrency;
concurrency.value = concurrency.max;
concurrency.oninput();
concurrency.disabled = false;
}
let rendering = null;
let start = null;
let interval = null;
let pool = null;
class State {
constructor(wasm) {
this.start = performance.now();
this.wasm = wasm;
this.running = true;
this.counter = 1;
this.interval = setInterval(() => this.updateTimer(), 100);
wasm.promise()
.then(() => {
this.updateTimer();
this.stop();
})
.catch(console.error);
}
updateTimer() {
const dur = performance.now() - this.start;
timingVal.innerText = `${dur}ms`;
this.counter += 1;
if (this.wasm && this.counter % 3 == 0)
this.wasm.requestUpdate();
}
stop() {
if (!this.running)
return;
console.timeEnd('render');
this.running = false;
pool = this.wasm.cancel(); // this frees `wasm`, returning the worker pool
this.wasm = null;
clearInterval(this.interval);
}
}
function render(scene) {
if (rendering) {
rendering.stop();
rendering = null;
}
rendering = new State(scene.render(concurrency.value, pool, ctx));
pool = null; // previous call took ownership of `pool`, zero it out here too
}

View File

@ -0,0 +1,364 @@
extern crate futures;
extern crate js_sys;
extern crate raytracer;
extern crate wasm_bindgen;
extern crate web_sys;
use std::cell::RefCell;
use std::cmp;
use std::rc::Rc;
use std::sync::atomic::{AtomicUsize, AtomicBool, Ordering::SeqCst};
use std::sync::atomic::ATOMIC_USIZE_INIT;
use std::sync::{Arc, Mutex, MutexGuard};
use futures::Future;
use futures::sync::oneshot;
use js_sys::{Promise, Error, WebAssembly, Uint8ClampedArray, Array};
use wasm_bindgen::JsCast;
use wasm_bindgen::prelude::*;
use web_sys::{CanvasRenderingContext2d, Worker, Event, ErrorEvent};
use web_sys::{DedicatedWorkerGlobalScope, MessageEvent};
macro_rules! console_log {
($($t:tt)*) => (log(&format_args!($($t)*).to_string()))
}
#[wasm_bindgen]
extern "C" {
#[wasm_bindgen(js_namespace = console)]
fn log(s: &str);
}
#[wasm_bindgen]
pub struct Scene {
inner: raytracer::scene::Scene,
}
static NEXT_ID: AtomicUsize = ATOMIC_USIZE_INIT;
#[wasm_bindgen]
impl Scene {
#[wasm_bindgen(constructor)]
pub fn new(object: &JsValue) -> Result<Scene, JsValue> {
console_error_panic_hook::set_once();
Ok(Scene {
inner: object.into_serde()
.map_err(|e| JsValue::from(e.to_string()))?,
})
}
pub fn render(
self,
concurrency: usize,
pool: WorkerPool,
ctx: CanvasRenderingContext2d,
) -> Result<RenderingScene, JsValue> {
let (tx, rx) = oneshot::channel();
let rx = rx.then(|_| Ok(JsValue::undefined()));
let data = Rc::new(RefCell::new(None::<Render>));
let pixels = (self.inner.width * self.inner.height) as usize;
let mut r = Render {
tx: Some(tx),
callback: None,
shared: Arc::new(Shared {
id: NEXT_ID.fetch_add(1, SeqCst),
need_update: AtomicBool::new(false),
scene: self.inner,
next_pixel: AtomicUsize::new(0),
remaining: AtomicUsize::new(concurrency),
rgb_data: Mutex::new(vec![0; 4 * pixels]),
}),
ctx,
};
let data2 = data.clone();
let callback = Closure::wrap(Box::new(move |msg: Event| -> Result<(), JsValue> {
let mut slot = data2.borrow_mut();
if let Some(mut data) = slot.take() {
match data.event(&msg) {
Ok(true) => {}
Ok(false) => *slot = Some(data),
Err(e) => {
*slot = Some(data);
return Err(e)
}
}
}
Ok(())
}) as Box<FnMut(_) -> Result<(), JsValue>>);
for worker in &pool.workers[..concurrency] {
let ptr_to_send = Arc::into_raw(r.shared.clone()) as u32;
let ptr_to_send = JsValue::from(ptr_to_send);
worker.post_message(&ptr_to_send)?;
worker.set_onmessage(Some(callback.as_ref().unchecked_ref()));
worker.set_onerror(Some(callback.as_ref().unchecked_ref()));
}
r.callback = Some(callback);
*data.borrow_mut() = Some(r);
Ok(RenderingScene {
inner: data,
promise: wasm_bindgen_futures::future_to_promise(rx),
pool,
})
}
}
#[wasm_bindgen]
pub struct WorkerPool {
workers: Vec<Worker>,
callback: Closure<FnMut(Event)>,
}
#[wasm_bindgen]
impl WorkerPool {
#[wasm_bindgen(constructor)]
pub fn new(max: u32) -> Result<WorkerPool, JsValue> {
let callback = Closure::wrap(Box::new(|event: Event| {
console_log!("unhandled event: {}", event.type_());
}) as Box<FnMut(Event)>);
let mut workers = Vec::new();
for _ in 0..max {
// TODO: what do do about `./worker.js`:
//
// * the path is only known by the bundler. How can we, as a
// library, know what's going on?
// * How do we not fetch a script N times? It internally then
// causes another script to get fetched N times...
let worker = Worker::new("./worker.js")?;
let array = js_sys::Array::new();
array.push(&wasm_bindgen::module());
// TODO: memory allocation error handling here is hard:
//
// * How to we make sure that our strong ref made it to a client
// thread?
// * Need to handle the `?` on `post_message` as well.
array.push(&wasm_bindgen::memory());
worker.post_message(array.as_ref())?;
worker.set_onmessage(Some(callback.as_ref().unchecked_ref()));
worker.set_onerror(Some(callback.as_ref().unchecked_ref()));
workers.push(worker);
}
Ok(WorkerPool {
workers,
callback,
})
}
}
impl Drop for WorkerPool {
fn drop(&mut self) {
for worker in self.workers.iter() {
worker.terminate();
}
}
}
#[wasm_bindgen]
pub struct RenderingScene {
inner: Rc<RefCell<Option<Render>>>,
promise: Promise,
pool: WorkerPool,
}
#[wasm_bindgen]
impl RenderingScene {
pub fn promise(&self) -> Promise {
self.promise.clone()
}
#[wasm_bindgen(js_name = requestUpdate)]
pub fn request_update(&self) {
if let Some(render) = self.inner.borrow().as_ref() {
render.shared.need_update.store(true, SeqCst);
}
}
pub fn cancel(self) -> WorkerPool {
if let Some(render) = self.inner.borrow_mut().take() {
// drain the rest of the pixels to cause all workers to cancel ASAP.
let pixels = render.shared.scene.width * render.shared.scene.height;
render.shared.next_pixel.fetch_add(pixels as usize, SeqCst);
}
for worker in self.pool.workers.iter() {
worker.set_onmessage(Some(&self.pool.callback.as_ref().unchecked_ref()));
worker.set_onerror(Some(&self.pool.callback.as_ref().unchecked_ref()));
}
self.pool
}
}
struct Render {
callback: Option<Closure<FnMut(Event) -> Result<(), JsValue>>>,
tx: Option<oneshot::Sender<()>>,
shared: Arc<Shared>,
ctx: CanvasRenderingContext2d,
}
struct Shared {
id: usize,
need_update: AtomicBool,
scene: raytracer::scene::Scene,
next_pixel: AtomicUsize,
remaining: AtomicUsize,
rgb_data: Mutex<Vec<u8>>,
}
#[wasm_bindgen]
extern {
type ImageData;
#[wasm_bindgen(constructor, catch)]
fn new(data: &Uint8ClampedArray, width: f64, height: f64) -> Result<ImageData, JsValue>;
}
impl Render {
fn event(&mut self, event: &Event) -> Result<bool, JsValue> {
if let Some(error) = event.dyn_ref::<ErrorEvent>() {
let msg = format!("error in worker: {}", error.message());
return Err(Error::new(&msg).into());
}
if let Some(msg) = event.dyn_ref::<MessageEvent>() {
let data = msg.data();
if let Some(data) = data.dyn_ref::<Array>() {
let id = data.pop();
let done = data.pop();
let image = data.pop();
if let Some(id) = id.as_f64() {
if id == self.shared.id as f64 {
self.ctx.put_image_data(image.unchecked_ref(), 0.0, 0.0)?;
return Ok(done.as_bool() == Some(true))
}
}
}
console_log!("unhandled message: {:?}", data);
return Ok(false)
}
console_log!("unhandled event: {}", event.type_());
Ok(false)
}
}
#[wasm_bindgen]
pub fn child_entry_point(ptr: u32) -> Result<(), JsValue> {
let ptr = unsafe {
Arc::from_raw(ptr as *const Shared)
};
assert_send(&ptr);
let global = js_sys::global()
.unchecked_into::<DedicatedWorkerGlobalScope>();
ptr.work(&global)?;
return Ok(());
fn assert_send<T: Send + 'static>(_: &T) {}
}
impl Shared {
fn work(&self, global: &DedicatedWorkerGlobalScope) -> Result<(), JsValue> {
// Once we're done raytracing a pixel we need to actually write its rgb
// value into the shared memory buffer for our image. This, however,
// requires synchronization with other threads (as currently
// implemented). To help amortize the cost of synchronization each
// thread processes a chunk of pixels at a time, and this number is how
// many pixes will be rendered synchronously before committing them back
// to memory.
const BLOCK: usize = 1024;
let width = self.scene.width as usize;
let height = self.scene.height as usize;
let end = width * height;
// Thread-local storage for our RGB data, commited back in one batch to
// the main image memory.
let mut local_rgb = [0; BLOCK * 4];
loop {
// First up, grab a block of pixels to render using an atomic add.
// If we're beyond the end then we're done!
let start = self.next_pixel.fetch_add(BLOCK, SeqCst);
if start >= end {
break
}
// Raytrace all our pixels synchronously, writing all the results
// into our local memory buffer.
let len = cmp::min(end, start + BLOCK) - start;
for (i, dst) in local_rgb.chunks_mut(4).enumerate().take(len) {
let x = (start + i) % width;
let y = (start + i) / width;
let ray = raytracer::Ray::create_prime(x as u32, y as u32, &self.scene);
let result = raytracer::cast_ray(&self.scene, &ray, 0).to_rgba();
dst[0] = result.data[0];
dst[1] = result.data[1];
dst[2] = result.data[2];
dst[3] = result.data[3];
}
// Ok, time to synchronize and commit this data back into the main
// image buffer for other threads and the main thread to see.
let mut data = self.rgb_data.lock().unwrap();
data[start * 4..(start + len) * 4]
.copy_from_slice(&mut local_rgb[..len * 4]);
// As a "nifty feature" we try to have a live progressive rendering.
// That means that we need to periodically send an `ImageData` to
// the main thread. Do so whenever the main thread requests it.
if self.need_update.swap(false, SeqCst) {
self.update_image(false, data, global)?;
}
}
// If we're the last thread out, be sure to update the main thread's
// image as this is the last chance we'll get!
if self.remaining.fetch_sub(1, SeqCst) == 1 {
let data = self.rgb_data.lock().unwrap();
self.update_image(true, data, global)?;
}
Ok(())
}
fn update_image(
&self,
done: bool,
data: MutexGuard<Vec<u8>>,
global: &DedicatedWorkerGlobalScope,
) -> Result<(), JsValue> {
// This is pretty icky. We can't create an `ImageData` backed by
// `SharedArrayBuffer`, so we need to copy the memory into a local
// JS array using `slice`. This means we can't use
// `web_sys::ImageData` right now but rather we have to use our own
// binding.
let mem = wasm_bindgen::memory()
.unchecked_into::<WebAssembly::Memory>();
let mem = Uint8ClampedArray::new(&mem.buffer())
.slice(
data.as_ptr() as u32,
data.as_ptr() as u32 + data.len() as u32,
);
drop(data); // unlock the lock, we've copied the data now
let data = ImageData::new(
&mem,
self.scene.width as f64,
self.scene.height as f64,
)?;
let arr = Array::new();
arr.push(data.as_ref());
arr.push(&JsValue::from(done));
arr.push(&JsValue::from(self.id as f64));
global.post_message(arr.as_ref())?;
Ok(())
}
}

View File

@ -0,0 +1,32 @@
// synchronously, using the browser, import out shim JS scripts
importScripts('raytrace_parallel.js');
let booted = false;
let lastPtr = null;
// Wait for the main thread to send us the shared module/memory. Once we've got
// it, initialize it all with the `wasm_bindgen` global we imported via
// `importScripts`.
//
// After our first message all subsequent messages are an entry point to run,
// so we just do that.
self.onmessage = function(args) {
self.onmessage = event => run(event.data);
const [module, memory] = args.data;
wasm_bindgen(module, memory)
.then(() => {
booted = true;
if (lastPtr)
run(lastPtr);
})
.catch(e => setTimeout(() => { throw e; })); // propagate to main `onerror`
};
function run(ptr) {
if (!booted) {
lastPtr = ptr;
return;
}
lastPtr = null;
wasm_bindgen.child_entry_point(ptr);
}